* remove check_psycopg() call from the setup.py, when installing from wheel it doesn't work anyway.
* call check_psycopg() function before process_arguments(), because the last one is trying to import psycopg and fails with the stacktrace, while the first one shows a nice human-readable error message.
* add psycopg2, psycopg2-binary, and psycopg3 extras, that will install psycopg2>=2.5.4, psycopg2-binary, or psycopg[binary]>=3.0.0 modules respectively.
* move check_psycopg() function to the __main__.py.
* introduce the new extra called `all`, it will allow to install all dependencies at once (except psycopg related).
* use the `build` module in order to create sdist bdist_wheel packages.
* update the documentation regarding psycopg and extras (dependencies).
Cache creates a lot of problems and prevents implementing a feature of automatic retention of physical replication slots for members with configurable retention policy.
Just read the entire cluster from Zookeeper instead and use watchers only for the `/leader` and `/config` keys.
There are cases when Citus wants to have a connection to the local postgres. By default it uses `localhost` for that, which is not alwasy available. To solve it we will set `citus.local_hostname` GUC to custom value, which is the same as Patroni uses to connect to Postgres.
1. Introduce DEBUG logs for callbacks
2. Configure log format in behave tests to include filename, line, and method name that triggered the callback and enable DEBUG logs for `patroni.postgresql.callback_executor` module.
P.S. unfortunately it works only starting from python 3.8, but it should be good enough for debug purpose because 3.7 is already EOL.
1. Enforce `_load_cluster()` after acquisition for the leader lock in ZooKeeper. Sometimes the notification from ZooKeeper was arriving too late and Patroni wasn't setting the `role=standby_leader`.
2. The `_get_node_to_follow()` method was falsely assuming that we still own the leader lock and returning the remote node instead of the new standby leader. While not a big issue per se, because the next HA loop usually fixes it, such behavior was causing flakiness of behave tests with Postgres 12 and older, where restart is required to update `primary_conninfo` GUC.
It used to work before #2652
Besides that fix a couple of more problems:
- make sure `_patch_or_create()` method isn't instantiating the `k8s_client.V1ConfigMap` object instead of `k8s_client.V1Endpoints` for non leader endpoints. The only reason it worked is that the JSON serialization for both object types is the same and doesn't include the object type name.
- `attempt_to_acquire_leader()` should immediately put the IP address of the primary to the leader endpoint. It didn't happen because of the oversight in the https://github.com/zalando/patroni/pull/1820.
- get rid from sleeps
- reduce retry_timeout
- avoid graceful Patroni shut down while DCS is "paused", just kill
Patroni and after that gracefully stop postgres
- don't try to delete Pod when Patroni is killed. If K8s API is paused it takes ages
The run time on my laptop is reduced from 2m to 1m28s.
Patroni is changing `synchronous_standby_names` and the `/sync` key in a very specific order, first we add nodes to `synchronous_standby_names` and only after, when they are recognized as synchronous they are added to the `/sync` key. When removing nodes the order is different: they are first removed from the `/sync` key and only after that from the `synchronous_standby_names`.
As a result Patroni expects that either actual synchronous nodes will match with the nodes listed in the `/sync` key or that new candidates to synchronous nodes will not match with nodes listed in the `/sync` key. In case if `synchronous_standby_names` was removed from the `postgresql.conf`, manually, or due the the bug (#2876), the state becomes inconsistent because of the wrong order of updates.
To solve inconsistent state we introduce additional checks and will update the `/sync` key with actual names of synchronous nodes (usually empty set).
If restarted in pause Patroni was discarding `synchronous_standby_names` from `postgresql.conf` because in the internal cache this values was set to `None`. As a result synchronous replication transitioned to a broken state, with no synchronous replicas according to the `synchronous_standby_names` and Patroni not selecting/setting the new synchronous replicas (another bug).
To solve the problem of broken initial state and to avoid similar issues with other GUC's we will read GUC's value if Patroni is joining running Postgres.
Patroni doesn't watch on all changes of member keys in order to not create too much load on ZooKeeper, but only subscribes to changes (ZNodes added or deleted) in the `/member` directory. Therefore when some important fields in the value are updated we remove and recreate ZNode in order to notify the leader or other members.
The leader should remove the member key only when the `checkpoint_after_promote` value is changed and replicas when the `state` is changed to/from `running`.
We don't care about the `version` field, because Patroni version can't be changed without restart, what will case ZooKeeper `session_id` to change it anyway.
This fix hopefully will reduce failures of behave tests on GH Actions.
Despite being validated by `IntValidator` some GUC's couldn't be casted directly to `int` because they include suffix. Example: `128MB`.
Close https://github.com/zalando/patroni/issues/2879
WARNING messages are produced by `urllib3` if Patroni is quickly restarted.
Instead we will check that the node is listen on a given port. This fact is actually enough to detect names clashes, while HTTP request could raise an exception is a few other cases, what might case false negatives.
Close https://github.com/zalando/patroni/issues/2881
Create permanent physical replication slots on standby nodes and use `pg_replication_slot_advance()` function to move them forward.
The `restart_lsn` is advanced based on values stored in the `/status` key by the primary node.
When slot is created on a replica it could be ahead the same slot on the primary and therefore there is some period of time when it doesn't protect WAL files from being recycled.
As was reported by @ants on Slack it could happen that `received_tli` is ahead of replayed timeline, therefore we should stop using it when deciding on pg_rewind if postgres is running and use only `IDENTIFY_SYSTEM` via replication connection.
It represents the `/status` key in DCS and makes it easier to introduce new values stored in the `/status` key without need to refactor all DCS implementations.
1. make _get_members_slots() method return data in the same format as _get_permanent_slots() method
2. move conflicting name handling from get_replication_slots() to _get_members_slots() method
3. enrich structure returned by get_replication_slots() with the LSN of permanent logical slots reported by primary
4. use the added information in the SlotsHandler instead of fetching it from the Cluster.slots
5. bugfix: don't try to advance logical slot that doesn't match required configuration
Due to a race condition Patroni was falsely assuming that the standby should be restarted because some recovery parameters (primary_conninfo or similar) were changed.
Close https://github.com/zalando/patroni/issues/2834
Sharing a single connection between REST API and the main thread (doing heartbeats) was working mostly fine, except when Postgres becomes so slow that REST API queries start blocking the main loop.
If the dedicated REST API connection isn't available we use the heartbeat connection as a fallback.
When running with the leader lock Patroni was just setting the `role` label to `master` and effectively `kubernetes.standby_leader_label_value` feature never worked.
Now it is fixed, but in order to not introduce breaking changes we just update default value of the `standby_leader_label_value` to the `master`.
1.0.3 removed support of `top_level` configuration parameter and builds
now are failing.
Besides that remove redundant pyyaml from requirements.docs.txt
Make it hold connection kwargs for local connections and all `NamedConnection` objects use them automatically.
Also get rid of redundant `ConfigHandler.local_connect_kwargs`.
On top of that we will introduce a dedicated connection for the REST API thread.
The same (almost) logic was used in three different places:
1. `Patroni` class
2. `Member` class
3. `_MemberStatus` class
Now they all inherit newly intoduced `Tags` class.
Cluster.get_replication_slots() didn't take into account that there can not be logical replication slots in a standby cluster replicas. It was only skipping logical slots for the standby_leader, but replicas were expecting that they will have to copy them over.
Also on replicas in a standby cluster these logical slots were falsely added to the `_replication_slots` dict.
1. stop using the same cursor all the time, it creates problems when not carefully used from different threads.
2. introduce query() method in the Connection class and make it return a result set when it is possible.
3. refactor most of the code that is relying (directly or indirectly) on the Connection object to use the query() method as much as possible.
This refactoring helps with reducing code complexity and will help with future introduction of a separate database connection for the REST API thread. The last one will help to improve reliability when system is under significant stress when simple monitoring queries are taking seconds to execute and the REST API starts blocking the main thread.
Due to historical reasons (not available before 9.6) we used `pg_current_wal_lsn()`/`pg_current_xlog_location()` functions to get current WAL LSN on the primary. But, this LSN is not necessarily synced to disk, and could be lost if the primary node crashed.
1. Unit tests should not really try accessing any resources.
2. Not doing so results in significant execution time of unit tests on Windows
In addition to that perform a request with timeout 3s. Usually this is more than enough to figure out whether resource is accessible.
Followup on #2724
Postgres supports two types of permissions:
1. owner only
2. group readable
By default the first one is used because it provides better security. But, sometimes people want to run a backup tool with the user that is different from postgres. In this case the second option becomes very useful. Unfortunately it didn't work correctly because Patroni was creating files with owner access only permissions.
This PR changes the behavior and permissions on files and directories that are created by Patroni will be calculated based on permissions of PGDATA. I.e., they will get group readable access when it is necessary.
Close#1899Close#1901
In issue #2735 it was discussed that there should be some warning around PostgreSQL parameters that do not pass validation.
This commit ensures something is logged for parameters that fail validation and therefore fall back to default values.
Close#2735Close#2740
Consider a following situation:
1. node1 is stressed so much that Patroni heart-beat can't run regularly and the leader lock expires.
2. node2 notice that there is no leader, gets the lock, promotes, and gets to a situation like it is described in 1.
3. Patroni on node1 finally wakes up, notice that Postgres is running as a primary, but without a leader lock and "happily" acquires the lock.
That is, node1 discarded promoting of node2, and the node2 after that it will not be possible to join the node2 back to the cluster, because pg_rewind is not possible when two nodes are on the same timeline.
To partially mitigate the problem we introduce an additional timeline check. If postgres is running as primary Patroni will consider it as a perfect candidate only if timeline isn't behind the last known cluster timeline recorder in the `/history` key. If postgres timeline is behind the cluster timeline postgres will be demoted to read-only. Further behavior would depend on `maximum_lag_on_failover` and `check_timeline` settings.
Since the `/history` key isn't updated instantly after promotion, there is still a short period of time when the issue could happen, but it seems that it is close to impossible to make it more reliable.
Close https://github.com/zalando/patroni/issues/2779
It was reported in #2779 that the primary was constantly logging messages like `Synchronous replication key updated by someone else`.
It happened after Patroni was stuck due to resource starvation. Key updates are performed using create_revision/mod_revision field, which value is taken from the internal cached. Hence, it is a clear symptom of stale cache.
Similar issues in K8s implementation were addressed by invalidating the cache and restarting watcher connections every time when update failed due to resource_version mismatch, so we do the same for Etcd3.
On Slack user reported that Patroni didn't run pg_rewind on one of the nodes after coming out of maintenance mode.
Steps that were executed:
0. The initial state: node1 - primary, node2 - replica
1. `patronictl pause`
2. On node2: pg_ctl promote
3. On node1: pg_ctl stop
4. Patroni on node1 notice that Postgres isn't running and removes the leader lock
5. Patroni on node2 notice that Postgres is running as a primary and takes the leader lock.
6. `patronictl resume`.
After that node1 started saying in logs:
`Waiting for checkpoint on node2 before rewind`.
Repeating this steps may not necessarily reproduce the problem, because presumably pg_rewind failed earlier on node1.
Such situation was possible because promote wasn't executed by Patroni and therefore `Rewind._state` wasn't explicitly reset and the code that ensures that CHECKPOINT after promote was run wasn't triggered.
As a mitigation following changes have been made:
1. retrigger pg_rewind checks after coming out of maintenance mode
2. run ensure CHECKPOINT after promote checks using `Rewind._state != REWIND_STATUS.CHECKPOINT` condition. It allowed to remove useless hook from `Postgresql.promote()`.
When we don't have permanent logical slots Patroni was updating the `/status` on every heart-beat loop even when LSN on the primary isn't moving forward.
The issue was introduced in the #2485
1. Take client certificates only from the `ctl` section. Motivation: sometimes there are server-only certificates that can't be used as client certificates. As a result neither Patroni not patronictl work correctly even if `--insecure` option is used.
2. Document that if `restapi.verify_client` is set to `required` then client certificates **must** be provided in the `ctl` section.
3. Add support for `ctl.authentication` and prefer to use it over `restapi.authentication`.
4. Silence annoying InsecureRequestWarning when `patronictl -k` is used, so that behavior becomes is similar to `curl -k`.
If one of backends crashed postmaster stops all backend and does crash recovery because shared memory could be corrupted. If something happens during this phase postgres may completely crash leaving pg_control with "in crash recovery" state.
Followup on #2726
To do that we use `pg_stat_get_wal_receiver()` function, which is available since 9.6. For older versions the `patronictl list` output and REST API responses remain as before.
In case if there is no wal receiver process we check if `restore_command` is set and show the state as `in archive recovery`.
Example of `patronictl list` output:
```bash
$ patronictl list
+ Cluster: batman -------------+---------+---------------------+----+-----------+
| Member | Host | Role | State | TL | Lag in MB |
+-------------+----------------+---------+---------------------+----+-----------+
| postgresql0 | 127.0.0.1:5432 | Leader | running | 12 | |
| postgresql1 | 127.0.0.1:5433 | Replica | in archive recovery | 12 | 0 |
+-------------+----------------+---------+---------------------+----+-----------+
$ patronictl list
+ Cluster: batman -------------+---------+-----------+----+-----------+
| Member | Host | Role | State | TL | Lag in MB |
+-------------+----------------+---------+-----------+----+-----------+
| postgresql0 | 127.0.0.1:5432 | Leader | running | 12 | |
| postgresql1 | 127.0.0.1:5433 | Replica | streaming | 12 | 0 |
+-------------+----------------+---------+-----------+----+-----------+
```
Example of REST API response:
```bash
$ curl -s localhost:8009 | jq .
{
"state": "running",
"postmaster_start_time": "2023-07-06 13:12:00.595118+02:00",
"role": "replica",
"server_version": 150003,
"xlog": {
"received_location": 335544480,
"replayed_location": 335544480,
"replayed_timestamp": null,
"paused": false
},
"timeline": 12,
"replication_state": "in archive recovery",
"dcs_last_seen": 1688642069,
"database_system_identifier": "7252327498286490579",
"patroni": {
"version": "3.0.3",
"scope": "batman"
}
}
$ curl -s localhost:8009 | jq .
{
"state": "running",
"postmaster_start_time": "2023-07-06 13:12:00.595118+02:00",
"role": "replica",
"server_version": 150003,
"xlog": {
"received_location": 335544816,
"replayed_location": 335544816,
"replayed_timestamp": null,
"paused": false
},
"timeline": 12,
"replication_state": "streaming",
"dcs_last_seen": 1688642089,
"database_system_identifier": "7252327498286490579",
"patroni": {
"version": "3.0.3",
"scope": "batman"
}
}
```
- make sure that physical replication slots are created even before the promote happened (when async executor is busy with promote).
- execute `txid_current()` with `synchronous_commit=off` so it doesn't accidentally wait for absent synchronous standbys when `synchronous_mode_strict` is enable and `synchronous_standby_names=*`. These standbys can't connect because replication slots weren't there.
- `synchronous_standby_names` wasn't set to `*` after bootstrap with `synchronous_mode` and `synchronous_mode_strict`.
- add `-c statement_timeout=0` to `PGOPTIONS` when executing `post_bootstrap` script.
Close https://github.com/zalando/patroni/issues/2738
If we know for sure that a few moments ago postgres was still running as a primary and we still have the leader lock and can successfully update it, in this case we can safely start postgres back not in recovery. That will allow to avoid bumping timeline without a reason and hopefully improve reliability because it will address issues similar to #2720.
In addition to that remove `if self.state_handler.is_starting()` check from the `recover()` method. This branch could never be reached because the `starting` state is handled earlier in the `_run_cycle()`. Besides that remove redundant `self._crash_recovery_executed`.
P.S. now we do not cover cases when Patroni was killed along with Postgres.
Lets consider that we just started Patroni, there is no leader, and `pg_controldata` reports `Database cluster state` as `shut down`. It feels logical to use `Latest checkpoint location` and `Latest checkpoint's TimeLineID` to do a usual leader race and start directly as a primary, but it could be totally wrong. The thing is that we run `postgres --single` if standby wasn't shut down cleanly before executing `pg_rewind`. As a result `Database cluster state` transition from `in archive recovery` to `shut down`, but if such a node becomes a leader the timeline must be increased.
Reorder some checks and verify that the old primary is already in the `running` state before checking replication. This check elliminates the race condition when replication started to work but node name is removed from the `synchronous_standby_names` because state isn't `running`.
implementations and terminologis are DCS specific:
- Etcd v2 calls is `quorum` read
- Etcd v3 calls it `linearizable` (vs `serializable`)
- Consul calls it `consistent`
Following DCS don't offer this feature:
- ZooKeeper calls it linearizable, but reads are sequentially consistent
- Raft - no quorum reads are possible ATM
- Kubernetes - uses Etcd under the hood, but provides no API to choose read consistency level
Close https://github.com/zalando/patroni/issues/1199
Setting it before calling the parent constructor didn't really work because it is being overwritten in the `etcd.Client.__init__()`.
The only viable way of doing it is passing a custom value to the parent class.
Close https://github.com/zalando/patroni/issues/2142
* Reduce flakiness of citus behave tests
- make a few attempts with timeout when checking registered nodes
- get rid from artificial sleep
- allow check_registration() function to check secondaries
These changes are useful for Quorum based failover (#2668) and future PR
that enhances Citus support by registering secondaries in `pg_dist_node`.
They somehow messed up with type hints what made pyright unhappy.
To solve it we explicitly pass the Group class to the group() decorator.
In addition to that bump pyright version.
now it checks that inside square brackets there is indeed IPv6.
In addition to that fix a little issue in the function itself so it returns exactly the same result as psycopg2.extensions.parse_dsn().
Close https://github.com/zalando/patroni/pull/2714
- the `_in_fligh` attribute is accessed from multiple threads and must be protected with mutex when it is changed
- allow adding tasks for `_in_fligh.group` from the `sync_pg_dist_node()` method when timeout is reached. Not doing so might result is indefinite transaction if REST API request from worker node failed.
Starting from 1.27 there is containerd process, which also uses k3s binary and being detected by pidof. Therefore we will search for "k3s server" string in the process list instead of just "k3s".
it helps to get rid of recurring patterns:
```python
retry.deadline = retry.stoptime - time.time()
if retry.deadline < XXX:
raise Exception(...) or return False
```
pass reference to a last known leader object in order to avoid obtaining it from the `AbstractDCS.cluster` cache.
This change is useful for Consul, Etcd3 and Zookeeper implementations.
It didn't took into account the fact that we can get a new lease after changing a TTL. In this case we have to update the exiting leader key with the new lease.
To solve it we introduce the on transaction 'failure' callback. The whole workflow looks like (schematically):
```python
txn(
compare=(old_value == self._name)),
success=put(self.leader_path, self._name, self._lease),
failure=txn(
compare=(create_revision == '0'),
success=put(self.leader_path, self._name, self._lease)
)
)
```
The problem was introduced in d98d6d0b02c9cc67464a7b1b31b1c5570c26e12d
- Always pass etcd3 key revision as a string
- Make sure the leader key isn't unconditionally overwritten. It may happen that the leader heart-beat loop didn't run properly and the session has expired. In this case the leader may create a new session and a new leader key. But, there are chances that the other node already created a leader key and we don't want to overwrite it.
- Try to sync HA loops between nodes by adding 0.5 seconds to timeout on non-leader nodes
Make it return the new `SyncState` object in order to avoid reading the new cluster state in the Ha.process_sync_replication().
Now it is a small optimization, but it will become very handy in the quorum commit feature.
- make parse_sync_standby_names() return NamedTuple instead of dict
- fix little issue in Postgresql.reset_cluster_info_state(), it should check global_config independently from cluster and cluster.config
- added pyrightconfig.json with typeCheckingMode=strict
- added type hints to all files except api.py
- added type stubs for dns, etcd, consul, kazoo, pysyncobj and other modules
- added type stubs for psycopg2 and urllib3 with some little fixes
- fixes most of the issues reported by pyright
- remaining issues will be addressed later, along with enabling CI linting task
Sometimes MacOS workers are so slow that Postgres shutdown might take more than 30s-40s, what breaks a test with replica reinit in parallel with the primary restart, because basebackup() does only two attempts and in a pause the replica remains running with empty PGDATA.
In addition to that increase timeouts in ignore_slots test.
Close#2637
1. make `SyncHandler.current_state()` return `CaseInsensitiveSet` instead of `list` objects.
2. take `sync_node_count` and `sync_node_maxlag` from the `Postgresql._global_config` instead of passing them as arguments.
3. Make `AbstractDCS.write_sync_state()` accept any `Collection`-like objects.
For now it implements:
- CaseInsensitiveDict()
- CaseInsensitiveSet()
Update `patroni.postgresql.sync.parse_sync_standby_names()` to use `CaseInsensitiveSet()` instead of `CaseInsensitiveDict()`
despite being EOL the number of downloads from pypi for 3.6 is on the first place, hence we need to support it.
Fixed a couple of problems in tests:
1. pytest complains about `call` imported from mock, mock.call() is fine
2. one test of Citus intergration was broken
Previously it used to compare between the leader and sync_standbys, while in some cases (actually most of them) the leader should be excluded.
This commit makes `matches()` method flexible:
1. The leader will be included to comparison only if requested
2. checks will be performed as case insensitive (like PG does)
Besides that, everywhere in code start using `cluster.sync.matches()` instead of `name in cluster.sync.members`.
It was possible to have it empty if the all cluster keys are missing in DCS. In this case the `Cluster` object was manually created with all values set to `None` or `[]` (including sync).
It already resulted in #2217, which is in fact wasn't a correct fix.
In order to solve it and reduce code duplication we introduce `Cluster.empty()` and `SyncState.empty()` methods, which will create corresponding empty objects and start using `Cluster.empty()` from all places where the empty `Cluster` object was manually created.
If communication with etcd nodes failed it is logical to start from scratch, from nodes that are listed in the config. But, it could happen that config is in fact outdated and all nodes in the real cluster were replaced.
Previously we used to track whether config file was changed, which turned out not to work in all possible cases.
The new strategy is a bit more different - if communication with all nodes failed we will continue keeping the last know topology and at the same time will try to figure out the new one by merging two lists together, the cached list and the list from the config file.
We made incorrect assumption that `citus_set_coordinator_host()` will trigger `pg_dist_node` sync. Instead we should also use `citus_update_node()` and call `citus_set_coordinator_host()` only during the bootstrap.
Adjust behave tests to verify that coordinator failover is visible on workers.
Configuration parameter is `kubernetes.retriable_http_codes` or `PATRONI_KUBERNETES_RETRIABLE_HTTP_CODES` environment variable.
These status codes are added to the default list of 500, 503, 504.
Close https://github.com/zalando/patroni/issues/2536
It could happen that Patroni is started up before PGDATA was mounted. In this case Patroni can't determine major Postgres version from PG_VERSION file. Later, when PGDATA is mounted, Patroni was trying to create the recovery.conf even if the actual Postgres major version is newver than 12.
To mitigate the problem we double check that the `Postgresql._major_version` is set before writing recovery configuration or starting postgres up.
Close https://github.com/zalando/patroni/issues/2434
Since a long time Patroni enforcing only one callback script running at a time. If the new callback is executed while the old one is still running, the old one is killed (including all child processes).
Such behavior is fine for all callbacks but on_reload, because the last one may accidentally cancel important ones, that for example updating DNS or assigning/removing Virtual IP.
To mitigate the problem we introduce a dedicated executor for on_reload callbacks, so that on_reload may only cancel another on_reload.
Ref: https://github.com/zalando/patroni/issues/2445
